Ensure to initialize |w_recalc| parameter.

Author: sekia

lib/Algorithm/LibLinear.pm

@@ -62,6 +62,7 @@ sub new {
         my $epsilon => +{ isa => Num, optional => 1, },
         my $loss_sensitivity => +{ isa => Num, default => 0.1, },
         my $nu => +{ isa => Num, default => 0.5, },
+        my $recalculate_weights => +{ isa => Bool, default => 0, },
         my $regularize_bias => +{ isa => Bool, default => 1, },
         my $solver => +{
             isa => $SolverDescriptor,
@@ -82,6 +83,7 @@ sub new {
         $loss_sensitivity,
         $nu,
         $regularize_bias,
+        $recalculate_weights,
     );
     bless +{
       bias => $bias,
@@ -170,6 +172,8 @@ sub is_regression_solver { $_[0]->training_parameter->is_regression_solver }
 
 sub loss_sensitivity { $_[0]->training_parameter->loss_sensitivity }
 
+sub recalculate_weights { $_[0]->training_parameter->recalculate_weights }
+
 sub training_parameter { $_[0]->{training_parameter} }
 
 sub train {
@@ -242,7 +246,7 @@ Current version is based on LIBLINEAR 2.48, released on January 5, 2025.
 
 =head1 METHODS
 
-=head2 new([bias => -1.0] [, cost => 1] [, epsilon => 0.1] [, loss_sensitivity => 0.1] [, nu => 0.5] [, regularize_bias => 1] [, solver => 'L2R_L2LOSS_SVC_DUAL'] [, weights => []])
+=head2 new([bias => -1.0] [, cost => 1] [, epsilon => undef] [, loss_sensitivity => 0.1] [, nu => 0.5] [, recalculate_weights => 0] [, regularize_bias => 1] [, solver => 'L2R_L2LOSS_SVC_DUAL'] [, weights => []])
 
 Constructor. You can set several named parameters:
 
@@ -272,6 +276,10 @@ Epsilon in loss function of SVR (C<-p>.)
 
 Nu parameter of one-class SVM (C<-n>.)
 
+=item recalculate_weights
+
+Whether to recalculate weights. This option is only valid for dual solvers for L2-regularized L1/L2-loss SVM (L2R_L1LOSS_SVC_DUAL / L2R_L2LOSS_SVC_DUAL).
+
 =item regularize_bias
 
 Whether to regularize the bias term (C<-R>, negated.)

src/liblinear.xs

@@ -25,6 +25,7 @@ alloc_parameter(pTHX_ int num_weights) {
     }
     parameter_->init_sol = NULL;
     parameter_->nr_weight = num_weights;
+    parameter_->w_recalc = false;
     return parameter_;
 }
 
@@ -394,7 +395,7 @@ MODULE = Algorithm::LibLinear  PACKAGE = Algorithm::LibLinear::TrainingParameter
 PROTOTYPES: DISABLE
 
 struct parameter *
-ll_new(klass, solver_type, epsilon, cost, weight_labels, weights, loss_sensitivity, nu, regularize_bias)
+ll_new(klass, solver_type, epsilon, cost, weight_labels, weights, loss_sensitivity, nu, regularize_bias, recalculate_weights = false)
     int solver_type;
     double epsilon;
     double cost;
@@ -403,6 +404,7 @@ ll_new(klass, solver_type, epsilon, cost, weight_labels, weights, loss_sensitivi
     double loss_sensitivity;
     double nu;
     bool regularize_bias;
+    bool recalculate_weights;
 CODE:
     int num_weights = av_len(weight_labels) + 1;
     if (av_len(weights) + 1 != num_weights) {
@@ -419,6 +421,7 @@ CODE:
     RETVAL->p = loss_sensitivity;
     RETVAL->nu = nu;
     RETVAL->regularize_bias = regularize_bias ? 1 : 0;
+    RETVAL->w_recalc = recalculate_weights;
     dXCPT;
     XCPT_TRY_START {
         int *weight_labels_ = RETVAL->weight_label;
@@ -518,6 +521,14 @@ CODE:
 OUTPUT:
     RETVAL
 
+bool
+ll_recalculate_weights(self)
+    struct parameter *self;
+CODE:
+    RETVAL = self->w_recalc;
+OUTPUT:
+    RETVAL
+
 int
 ll_solver_type(self)
     struct parameter *self;

t/Algorithm/LibLinear/TrainingParameter.t

@@ -16,6 +16,7 @@ my $tolerance = 10e-15;
             cost => num(1, $tolerance),
             epsilon => num(0.1, $tolerance),
             is_regression_solver => bool(0),
+            recalculate_weights => bool(0),
             weights => [],
         ],
     );
@@ -47,6 +48,22 @@ my $tolerance = 10e-15;
     );
 }
 
+{
+    my $learner = new_ok 'Algorithm::LibLinear' => [
+        recalculate_weights => 1,
+    ];
+    cmp_methods(
+        $learner,
+        [
+            cost => num(1, $tolerance),
+            epsilon => num(0.1, $tolerance),
+            is_regression_solver => bool(0),
+            recalculate_weights => bool(1),
+            weights => [],
+        ],
+    );
+}
+
 my $data_set = Algorithm::LibLinear::DataSet->new(data_set => [
     +{ feature => +{}, label => 1 },
 ]);
@@ -61,6 +78,14 @@ my @cases = (
       ],
       error_pattern => qr/p < 0/,
     },
+    +{
+      constructor_params => [
+        recalculate_weights => 1,
+        # Arbitrary value except L2R_L2LOSS_SVC_DUAL / L2R_L1LOSS_SVC_DUAL.
+        solver => 'L2R_L2LOSS_SVR',
+      ],
+      error_pattern => qr!Recalculating w in the end is only for dual solvers for L2-regularized L1/L2-loss SVM!,
+    },
 );
 for my $case (@cases) {
     my $learner = Algorithm::LibLinear->new(@{ $case->{constructor_params} });